Two popular open-source models - Google's Gemma 4 31B and Alibaba's Qwen3.5 27B - were put head-to-head on SVG generation, with results showing meaningful stylistic differences between the two.
The test used Q4 quantized versions of both models from Unsloth. Quantization is a compression technique that reduces a model's file size and memory requirements by storing numbers at lower precision - Q4 means each weight uses 4 bits instead of the standard 16 or 32. The tradeoff is a small quality drop in exchange for running on consumer-grade hardware rather than expensive server GPUs. Unsloth is a popular tool in the local AI community for producing efficient quantized model versions.
SVG (Scalable Vector Graphics) is a code-based image format. Instead of pixels, SVGs describe shapes, paths, and colors using XML markup, which means generating a valid, attractive SVG requires a model to handle both visual reasoning and precise syntax at the same time. It's a harder task than it looks.
Even at similar parameter counts - 31B vs 27B - the two models produce noticeably different aesthetics. That gap matters if you're building tools that output diagrams, icons, or data visualizations through AI-generated code rather than raster images.
Both models are freely available for local deployment, making the Unsloth quantized versions a practical starting point for anyone who wants SVG generation without paying per-API-call.